1. Mohammed Abouzour, Kenneth Salem, Peter Bumbulis Presentation by Mohammed Abouzour SMDB2010

2.  Database servers have two main architectures:  Worker-per-connection: (Oracle, Sybase ASE) A worker is dedicated for each connection for the life of the connection  Worker-per-request: (Oracle, Microsoft SQL Server, Sybase ASE, Sybase SQL Anywhere) A worker pool and a request queue: each worker picks and complete one request at a time. No guarantee that the same worker will service the same connection. 2

3.  Efficient in resource use and handling of large number of connections (1)  How to choose the size of the worker pool?  A large worker pool: Increases the concurrency level of the server Increases contention on server resources Increases working set size of server  A smaller worker pool: Under utilization of hardware resources Limit concurrency level of workload Possibility of a server hang due to no workers available to handle outstanding requests 3

4.  Dynamically adjust the size of the worker pool (the multiprogramming level of the server)  Benefits of dynamic MPL:  One less parameter for DBAs to worry about  One of the design goals of Sybase SQL Anywhere  Improve server performance for different workloads  Better handling of changes in workload transaction mix  We used throughput as a way to detect degraded server performance 4

5.  Introduction to Problem  Controller Design and Tuning Algorithms  Experimental Study  Results  Future Work 5

6. 6

7. 7

8.  Tuning algorithms are based on a paper by Heiss and Wagner (2) :  Incremental Steps algorithm (IS)  Parabola Approximation (PA)  Original algorithms were studied in simulation  Our contribution to the original algorithms (3) :  We have extended the IS algorithm  We have used a commercial database system and an industry standard benchmark (TPC-C)  We have studied the behaviour of the above two algorithms under different workloads 8

9. NotationDescription titi End of current control interval and start of the next control interval t i+1 P(t i )The actual measured throughput level of the server at time t i. Throughput is the number of request completed during the previous measurement interval n(t i )The workload concurrency level at time t i n * (t i )The number of workers the server has available. 9

10.  Original IS algorithm did not handle a throughput curve that flattens out  Modified the IS algorithm as follows:  Follow the throughput curve up as long as percentage of change in throughput is greater than C times the percentage of change in number of threads  Follow the throughput curve down as long as throughput increases Built-in bias to keep the number of threads as low as possible 10

11.  Model the throughput curve as a parabolic function:  Collect two points from the last two measurements and use the origin  Only make adjustments if the computed parabolic curve is concave down, otherwise just make a random adjustment 11

12.  Used the TPC-C workload for our benchmark  Used two workloads variations:  TPCC1:Standard TPC-C transaction mix  TPCC2:Modified TPC-C transaction mix (lighter on I/O and higher CPU component)  Two different buffer pool configurations:  BIGMEM:1.0GB  SMALLMEM:0.5GB  Three different workload configurations:  TPCC1-BIGMEM  TPCC1-SMALLMEM  TPCC2-BIGMEM 12

13.  Workload characterization experiments:  Ran each workload under different fixed MPL setting and generated a throughput curve for each workload  Experiments:  Auto-tuning experiments: We ran each of the workloads and allowed the auto tuning algorithms to adjust MPL dynamically  Changing workloads experiments: We start with one workload and half way in the experiment we switch to a different workload 13

14. TPCC1-BIGMEM TPCC1-SMALLMEM 14

15. Hill Climbing Global Parabola Approximation 15

16. Hill Climbing Global Parabola Approximation 16

17. Hill Climbing Global Parabola Approximation 17

18.  Not all workloads in database systems have hill- shaped throughput curve  Both algorithms were able to react appropriately to workloads changes  The Hill Climbing algorithm:  Smoother in making adjustments (more desirable)  Takes longer to reach optimal values  Can get stuck in a local minima  The Global Parabola approximation algorithm:  Takes variable step sizes and hence reacts faster to changes  Very sensitive to normal minor workload fluctuations 18

19.  A hybrid controller:  Uses the Hill Climbing algorithm for a number of steps and collects data  Uses the data collected with the Parabola Approximation algorithm  We reduced the control interval length  Hill Climbing modifications:  Take dynamic step sizes to speed up algorithm  C parameter reduced to make algorithm more aggressive  Parabola Approximation modifications:  Use least-square method to do curve fitting on collected data  Results:  90-95% of optimal tpmC compared to a hand tuned version  In other internal CPU-bound benchmark was able to improve performance by 32% 19

20.  Dynamically adjusting the MPL:  Improves server performance in some workloads compared to a fixed MPL  Handles changes to workload conditions effectively  Brings us one step closer to a self managing database system  Algorithm parameters can affect the performance of the algorithms:  Control interval length  Hill Climbing: Step sizes C parameter  The algorithms did not behave good for the following workloads:  Bursty workloads: busy periods followed by idle periods followed by busy periods  Long running requests: e.g. creating an index Use response time to help the auto tuning algorithm 20

21. 1. B. Schroeder, M. Harchol-Balter, A. Iyengar, E. Nahum, and A. Wierman, “How to determine a good multi-programming level for external scheduling,” in Proceedings of the 22 nd International Conference on Data Engineering. Washington, DC, USA: IEEE Computer Society, 2006, p. 60. 2. H.-U. Heiss and R. Wagner. “Adaptive load control in transaction processing systems,” in VLDB ‘91: Proceedings of the 17 th International Conference on Very Large Data Bases. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc., 1991, pp. 47-54. 3. M. Abouzour, “Automatically tuning database server multiprogramming level,” Master’s thesis, University of Waterloo, 2007 21

22. 22

23.  150 Warehouses  10 clients per warehouse  500 connections  C = 0.4, Delta = 10  Gamma=0.333, Delta = 10  Windows 2003 32-bit  Dual 1.80Ghz Intel Xeon  48 Spindles 10K RPM SCSI drives 23